Valuing health using visual analogue scales and rank data: does the visual analogue scale contain cardinal information?

Transcription

1 Academic Unit of Health Economics LEEDS INSTITUTE OF HEALTH SCIENCES Working Paper Series No. 09_01 Valuing health using visual analogue scales and rank data: does the visual analogue scale contain cardinal information? Hulme, C. a, * and Edlin, R. a a Academic Unit of Health Economics, University of Leeds, Leeds, UK * Dr Claire Hulme Academic Unit of Health Economics Leeds Institute of Health Sciences University of Leeds 101 Clarendon Road LEEDS LS2 9LJ Tel: +44 (0) c.t.hulme@leeds.ac.uk Disclaimer The series enables staff and student researchers based at or affiliated with the AUHE to make recent work and work in progress available to a wider audience. The work and ideas reported here do not represent the final position and need to be treated as work in progress. The material and views expressed in the series are solely those of the authors and should not be quoted without their permission.

2 2 Abstract Valuation studies have favoured standard gamble (SG) and time trade off techniques (TTO) over visual analogue scales (VAS). The lack of observable trade off properties of VAS precludes preferences being measured on a cardinal scale. The inferences, and indeed the premise, of many perceived VAS deficiencies have been debated and, as part of this on-going debate, Brazier and McCabe (2007) asked whether VAS data added anything to rank data; echoing the suggestion that VAS functioned primarily as a prop for ranking exercises (Torrance et al 2001). Previous studies suggest that analyses of ordinal (rank data) can provide valuation functions broadly equivalent to cardinal health state data models (Salomon, 2003; McCabe et al, 2006). This paper adds to the debate by considering whether ordinal preferences, cardinal differences, and cardinal scores from VAS data provide substantively different valuation algorithms. That is, is potentially valuable information lost by using ordinal (rank) data rather than cardinal VAS scores? In the case where a cardinal difference model can provide a substantively better fit than an ordinal preference model, then we can say that the VAS contains useful cardinal information that cannot be incorporated into an ordinal model. A further aim is to assess whether VAS-based ordinal preferences (rather than TTO- or SG-based ordinal preferences) are likely to be sufficient reliably to inform policy. Using rescaled data from a UK general population survey seven separate random effects logit regressions were carried out for ordinal preferences and seven for cardinal differences. In contrast to previous studies the analyses found that ordinal preference models appear to give different results to cardinal data. Ordinal preference data performed worse than the cardinal difference suggesting that: (1) VAS contains at least some relevant and useful cardinal data and (2) ignoring such data worsens performance of the resulting measures. Keywords: Visual analogue scales; rating scale; ordinal preferences; health state valuation; preference-based health measures

3 3 Introduction The role of visual analogue scales (VAS) in valuation studies has been extensively discussed. Valuation studies have favoured standard gamble (SG) and time trade off techniques (TTO) over VAS with proponents of the former techniques arguing that the lack of theoretical foundation means that valuations using VAS cannot relate to the underlying theory of QALYs (Johannesson et al, 1996). The lack of observable trade off properties precludes preferences being measured on a cardinal scale. VAS data have also been criticised for the absence of uncertainty and problems with bias due to framing effects including context bias and end state aversion (Parkin and Devlin, 2006; Torrance et al, 2001). Torrance et al also highlight issues associated with the lack of agreement between VAS and SG results and anchoring, in particular definitions such as best imaginable and worst imaginable state. The inferences, and indeed the premise, of many perceived VAS deficiencies have been debated (Torrance et al, 2001; Parkin and Devlin, 2006; Brazier and McCabe, 2007) and, as part of this ongoing debate, recently Brazier and McCabe (2007) asked whether VAS data added anything to rank data; echoing Torrance and colleagues suggestion that VAS functioned primarily as a prop for ranking exercises. Previous studies suggest that analyses of ordinal (rank data) can provide valuation functions broadly equivalent to cardinal health state data models (Salomon, 2003; McCabe et al, 2006). Using a conditional logit model, Salomon proposes estimation of cardinal values be carried out using aggregate data on ordinal rankings. The model uses TTO data from a UK general population survey, The Measurement and Valuation of Health (MVH) (Dolan et al, 1994). Within the survey health states were described using the EQ-5D descriptive system (five dimensions: mobility, self care, usual activities, pain/discomfort, anxiety/depression) and respondents asked to describe their own health state before ranking 13 hypothetical states (taken from 42) plus dead and unconscious. Rankings were followed by ratings of the same states using a VAS. Finally a series of TTO questions were used for the 13 states. Respondents initially indicated whether each state was preferred to dead. When considering a state j, the model assumes that respondent i will have a latent utility Uij composed of a systematic component value ( j ) and an error term ( ij ). When comparing two states, it is assumed that the state with the higher latent utility estimate will be preferred. If each health state has the same underlying value across all respondents, then State j will be preferred over

4 4 State k where the difference in error terms ( ) is less than the difference in underlying values ( j k ), and State k where not. ij ik The individual-level rank data are treated as a series of choices. Initially one state, with the best rank, is chosen over all others, then the state with the second rank over all others with the exception of the first and so on. This is equivalent to a ranking of n states and provides n(n-1)/2 discrete preferences. The model assumes that the expected value of the latent utility of each health state is a linear function of the ratings of the EQ-5D domains j rescaled in three distinct ways: θ. Before running the regression the data are x j 1. normalisation to match the scale of observed TTO values in the data 2. normalisation to produce a utility of 0 for the state 3. normalisation to produce a utility of 0 for dead Additional analyses run the rankings in reverse order to take account of a right skewed distribution in order to consider whether this produces important differences. Salomon found all three rescaling options predictions were strongly correlated with the observed TTO values (Pearson s r for Options 1 and 2 and for Option 3). Using ICC, normalisation to match the scale of observed TTO values in the data was the best fit and similar to that of the TTO (0.974 vs ). Comparison of modelled and observed TTO values by state show that the four states with the largest discrepancies between predicted values from the rank model and mean observed TTO values, all included Level 3 on the dimension of pain, and all were states with the largest differences in rank positions between the direct ordering exercise and the TTO. Comparison of inverted and non-inverted rank orderings showed comparable results and when compared to the observed TTO values were almost identical to the main model. These results led Salomon to the conclusion that the information content of aggregate rank data is similar to that of the TTO. However, Salomon highlights a number of important limitations or areas of further research including whether utilities are correlated across health states at the level of the individual (the assumption of independence from irrelevant alternatives) and choice of scale anchors which is complicated by differences in the relative ranking of dead between the two methods (unlike the SG and TTO being dead is not a health state for VAS which causes confusion for respondents (Brazier and McCabe, 2007).

5 5 Building on Salomon s model, McCabe et al (2006) estimate conditional logistic regression models for HUI2 and SF-6D using ordinal preference data to explore whether the results are comparable to those estimated by SG. Both the HUI2 and SF-6D have six dimensions (HUI2 sensation, mobility, emotion, cognition, self-care and pain; SF-6D physical functioning, role limitations, social functioning, pain, mental health and vitality). HUI2 respondents ranked eight health states from its classification plus full health and dead (McCabe et al 2005a; McCabe et al 2005b) and then valued the same states using SG. For the SF-6D a representative sample of the UK population were asked to value a sample of SF-6D health states using SG. Respondents were asked to rank five health states plus the best and worse states defined by the SF-6D. Using SG techniques respondents valued each one of five health states with against the best and worst SF-6D health states as alternative outcomes and then valued the worse measurable SF-6D (the pits state) in relation to dead. A sixth SG valuation was dependent on whether respondents had value PITS higher or lower than dead. McCabe et al s model, like Salomon s, assumes independence of irrelevant alternative (ranking of the pair not affected by other states ranked in the same exercise). The model assumes that respondent i will have a latent utility Uij when considering a state j, now comprising a systematic component value ( j ) and an error term ( ij ) and a dummy variable for dead (D). Again, respondents are assumed to choose the option with greater latent utility, and the expected value of the unobserved utility is a linear function of categorical level on the dimensions of each dataset: j x j θ D, where x j is composed of dummy variables for all dimension-level combinations (top score items appear as the baseline). All coefficients are rescaled by 1 to provide predictions at dead (0) and full health (1). In the case of the HUI2 the results show similarity between rank and SG models; although the former has one more inconsistency and doesn t distinguish as clearly between levels of mobility. For the SF- 6D model the rank model differs to that of the SG. Despite the rank model having a lower number of inconsistencies the predictive performance is better in the SG model. McCabe et al also note that there is evidence in the models that the assumption of independence doesn t hold that models are sensitive to exclusion of states ranked highly or lowly. They conclude that the rank models performed better than expected given the different informational content and that the results are consistent with the existence of a latent utility function.

6 6 The two studies outlined show promise for the use of rank data in valuing health. Whilst previous papers use TTO and SG, it is arguably inefficient to use intensive survey methods in order to obtain rank data only. If rank data are sufficient to create a valuation algorithm, then VAS values may be an appropriate way to obtain this data. Parkin and Devlin (2006) argue that VAS may provide some advantages over more complex methods in areas of feasibility, reliability and practicality (in particular, by allowing postal surveys). This paper uses the same dataset as Salomon but concentrates on VAS data. We consider whether the information contained in the VAS rank data provides similar results to the cardinal VAS data. That is, do we lose potentially valuable information by using ordinal (rank) data rather than cardinal VAS scores? The ordinal data infers a preference between two states, which provides different information than the VAS score the ordinal data considers the difference between states but does not consider the intensity of preference between them, whilst the cardinal score does not (in itself) consider differences between states but provides a level of utility. More useful comparisons may be between ordinal preferences (between states) and cardinal differences (between states), and cardinal differences (between states) and the VAS scores (level only). These cardinal differences have been considered previously. Dolan and Roberts (2002) investigated the cardinal preferences between states and explored whether a tariff with better predictive ability can be calculated using differences between EQ5D states and rather than the levels themselves. Roberts and Dolan (2004) also consider the degree to which individual ordinal preferences (inferred from TTO responses) correspond to differences in the mean TTO levels across individuals. They suggest a large degree of heterogeneity in preferences between individuals; this necessitates the use of random effects within this paper. This paper considers whether ordinal preferences, cardinal differences, and cardinal scores from VAS data provide substantively different valuation algorithms. In the case where a cardinal difference model can provide a substantively better fit than an ordinal preference model, then we can say that the VAS contains useful cardinal information that cannot be incorporated into an ordinal model. A further aim is to assess whether VAS -based ordinal preferences (rather than TTO- or SG-based ordinal preferences) are likely to be sufficient reliably to inform policy.

7 7 Data and preparation of dependent values Whilst the data from the UK general population survey (Dolan et al, 1994) is outlined in brief above, in more detail 3395 interviews were conducted; the sample was representative of the general population. Each interview consisted of five components. The first was self reported health using the EQ-5D descriptive system and VAS; secondly a ranking exercise consisting of 15 pre-determined states in which each state (including full health and dead) lasts for 10 years, with the dead state after 10 years. This was followed by VAS rating of the same states on a 0 to 100 scale. The next component was a TTO exercise in which full health and dead were removed (leaving just 13 states). VAS scores are used here without applying a correction for range frequency bias (Parducci and Weddell, 1986). Schwartz (1998) found that some VAS data appeared to provide cardinal values only when corrected for this bias. We considered applying the Parducci-Weddell correction (assuming an equal contribution to underlying value and spreading factors, w = 0.5) but found that this would have required omitting nearly half our data. Here, 30.5% of individual observations were above the maximum figure feasible under rank-frequency bias and a further 12.3% of observations are below the minimum feasible figure. For each respondent, individual VAS scores were first rescaled with dead and full health at 0 and 1. Those states rated worse than dead were rescaled so that the worse conceivable health state (i.e. 0 on the original VAS scale) would receive the value -1. (In the analysis of the TTO data, a similar assumption was made to limit the lowest value received to -1.) Three variables were defined as dependent variables for the subsequent analysis. When considering comparisons of states, the ORD and CARD variables are used. Here, the ORD variable takes the value 1 if the first state in a comparison is preferred, 0 if the second state is preferred, and 0.5 given indifference between them. Details of the assumptions used to form comparisons appear in the following section. The CARD variable presents the difference in (scaled) cardinal values between the states. The LEVEL variable considers the disutility from full health, and is used within the regressions to predict cardinal levels. When reporting coefficients, all results from the cardinal level regressions are multiplied by -1 in order to provide a measure of utility and hence greater comparability to the other results.

8 8 Analysis The existing literature uses a variety of methods, and some methods may be more or less appropriate to particular types of data than others. As such, the general aim is to identify an appropriate valuation algorithm for each type of data (ordinal preference, cardinal difference, VAS score) and compare their respective predictive abilities against mean EQ-5D VAS scores. Seven separate random effects logit regressions were carried out for different datasets representing ordinal preferences over dead, full health and the 13 states valued by each individual. Seven similar regressions were run for the cardinal differences between states using a standard random effects model. In both cases the random effect identifies individual respondents. All datasets were first prepared by scaling individual VAS responses against the values provided for dead (to 0) and full health (to 1). Values worse than dead were restricted to fall on the interval (0, -1] using the lower anchor of the VAS scale. The following groups were excluded: Those who failed to give a value for dead, full health, or the pits state (33333) Valued fewer than three other states in addition to these three Rated dead as better than full health, own health as worse than pits, or pits as better than full health. Rated fewer than three states better than dead. Of the 3395 individuals surveyed, 162 were excluded from the full MVH dataset, in comparison to the 84 exclusions in the original study. The seven datasets used in the regressions were constructed as follows. Pairwise combinations of all states: Exhaustive Following Salomon s methodology, our exhaustive dataset assumes independence of irrelevant alternatives between any two states. Here, the most preferred state in the ordinal ranking of n states is compared against the (n-1) states it is preferred to, the next state in our ordered list is compared against the (n-2) states (i.e. excluding the most preferred state), the next against (n-3) states and so on. Here the ranking of n states provides n(n-1)/2 preferences; for an individual providing 15 valuations, this leads to 105 distinct ordinal preferences and cardinal differences.

9 9 Extractions from a rank order: Proximate, Random 1, Random 2 Three datasets were prepared from a ranking between states. The proximate dataset uses the method from McCabe et al, and utilises each individual s orderings of the states. Here, the most preferred state was compared to the next most favoured state; this latter state to the one following, and so on. Each state typically appears on two occasions for an individual once against the next most preferred state, and once against the next least preferred state; the most and least preferred states only appear once. The rank order of n states provides n-1 preferences; for an individual providing 15 valuations, this leads to 14 distinct ordinal preferences and cardinal differences. In the Random 1 and Random 2 datasets a similar process was used but the states were not ordered by preference. Instead, a random order was generated between the states and the methodology above repeated to again provide n-1 distinct ordinal preferences and cardinal difference where n states were valued. Comparisons against a defined state: vs Full, vs Dead, vs Pits When constructing a regression based on cardinal differences Dolan and Roberts (2002) computed preferences against the EQ-5D state In theory, any state could be used for this purpose, and we consider full health, dead and pits here. In each case, preferences for n states will provide (n-1) ordinal preferences and cardinal differences. For each of these datasets, dummies were created that reflect the differences between two states in a comparison. Dead is coded as for full health with the exception of the values on the variable D (where dead is present as a first or second state, see Table 1 below). For example, suppose that the first state considered is full health, EQ-5D state and the second state considered is in which there are no problems in any domain with the exception of usual activities where the respondent indicates I am unable to perform my usual activities. Since the states are identical in all dimensions except usual activities, MO21, SC21, PD21, AD21, MO32, SC32, PD32, and AD32 all take the value 0. UA21 and UA32 both take the value 1, reflecting the higher health in the first state. D = 0 because neither state involves dead and N3 = -1 because only the second state involves a dimension at Level 3.

10 10 Table 1: Variable definitions Name Levels Coding MO21 Difference in mobility: levels 2 and 1 1 first state is level 1, SC21 Difference in self care: levels 2 and 1 UA21 Difference in usual activities: levels 2 and 1 PD21 Difference in pain/discomfort: levels 2 and 1 AD21 Difference in anxiety/depression: levels 2 and 1 second state level 2 or 3-1 first state is level 2 or 3, second state is level 1 0 all other cases MO32 Difference in mobility: levels 3 and 2 1 first state is level 1 or 2, SC32 Difference in self care: levels 3 and 2 UA32 Difference in usual activities: levels 3 and 2 PD32 Difference in pain/discomfort: levels 3 and 2 AD32 Difference in anxiety/depression: levels 3 and 2 second state level 3-1 first state is level 3, second state is level 1 or 2 0 all other cases D N3 Dead present as first or second state Is an EQ-5D 3 rd level present 1 state 1 is dead -1 state 2 is dead 0 dead not present in either state 1 first state has an EQ-5D level 3, second state has no EQ-5D level 3-1 first state has no EQ-5D level 3, second state has an EQ-5D level 3 0 all other cases Prior to all regressions except that involving LEVEL, data was randomly switched to eliminate a constant term in the cardinal differences regression. A uniform distribution on (0, 1) was estimated and a variable defined as to whether the value was above or below 0.5. If below 0.5, all the variables in the above table and CARD were multiplied by -1; in essence, the first and second states in any comparison were reversed. In these cases, ORD was recoded as 1 ORD. Where items were found to have the wrong sign, they were removed from the regression (in order of least significance to most significant). Results Table 2 displays the coefficients from the fifteen regressions. Bolded figures within the table indicate significance at 5%, with removed items left blank. The regressions differ in the number of insignificant and omitted items. In the regression on cardinal variable LEVEL all the EQ-5D variables in Table 1 above are significant with the exception of the N3 term; a similar pattern is observed in the cardinal Random 1, Random 2, and vs Dead models. All variables are significant in the exhaustive logit model

11 11 considering ordinal preferences including the N3 term. At the opposite extreme, the model considering ordinal preferences against the Dead state has only five EQ-5D significant variables with the correct sign (plus the constant): MO32, SC32, UA21, PD32, and AD32. In general, worsening health from Level 1 to Level 2 is unlikely to be sufficient for a health state to shift from better-thandead to worse-than-dead; the regression contains only one of five possible 2 1 shifts. Here, the cardinal model definitely outperforms the ordinal model. A fairer reflection of performance, however, is to consider the accuracy of a range of methodologies. Table 2: Regression coefficients from random effects and random effects logit models Exhaustive Logit Ranking: Preferences: Proximate Random 1 Random 2 vs FH vs Dead vs Pits CARDINAL LEVEL CARD CARD CARD CARD CARD CARD CARD MO MO SC SC UA UA PD PD AD AD D N Constant ORDINAL ORD ORD ORD ORD ORD ORD ORD MO MO SC SC UA UA PD PD AD AD D N Constant Bolded figures are significant at 5%. In order to compare the regression results against mean EQ-5D VAS scores, we estimate predictions for the 42 states valued in the MVH study and rescale these to the standard scale. (Unconscious was removed as there is no classification for this state on the EQ-5D.) Table 3 presents a summary of the cardinal difference and ordinal preference model fits, including their range and errors against the mean VAS scores. The range of the unscaled VAS predictions varies greatly, from for the cardinal Proximate model to -22 to 27 in the ordinal vs Full Health model. Scaled, the ordinal models appear typically place dead as the worst or near the worst

12 12 possible state (typically valued at around 0.00), whilst in the cardinal models the worst state is typically valued between -0.3 and Of the cardinal differences models, it appears that the vs Full Health model provides the lowest root mean squared error (RMSE) at 0.14 and also fewest absolute errors above 0.05, 0.10 and 0.20 (at 57%, 32% and 7% of cases). This model is taken to be our preferred cardinal difference model. For the ordinal difference models, it is clear that they have a generally higher RMSE (around 0.2 to 0.3 rather than ) and that three models appear to have a better performance than the others. Here, the Exhaustive, Random 1 and Random 2 models all have an RMSE around In all three cases, almost all cases have an absolute error above 0.05 and approximately half of the cases have an absolute error above We use the Exhaustive model as our preferred ordinal preference model but do so with caveats, as detailed in the discussion below. The performance of both the cardinal difference and ordinal preference models can be compared to that of the cardinal regression predicting LEVEL. Here, in contrast to other models, a lower proportion of predictions had absolute errors below 0.05 (50%), 0.10 (23%) and 0.20 (2%). Figure 1 presents the cumulative proportion of states by absolute error, where the higher values indicate superior fit. It is clear here that the VAS levels regression generally dominates, with the preferred cardinal difference model outperforming the preferred ordinal preference model. Table 3: Fit of the cardinal difference and ordinal preference models Exhaustive Ranking: Preferences vs: Logit Proximate Random 1 Random 2 Full health Death Pits CARDINAL (SCALED) DIFFERENCES MAX MIN ERRORS MAX RMSE by state %errors above % 84% 73% 73% 57% 75% 73% %errors above % 73% 45% 52% 32% 50% 48% %errors above % 59% 18% 18% 7% 18% 18% ORDINAL (SCALED) PREFERENCES MAX MIN ERRORS MAX RMSE by state

13 13 %errors above % 95% 93% 95% 95% 95% 95% %errors above % 91% 86% 84% 93% 95% 95% %errors above % 66% 48% 48% 64% 68% 84% Figure 1: Fit of the preferred models 100% 90% 80% 70% Cumulative Fit 60% 50% 40% VAS Levels Regression Cardinal (vs Full Health) Ordinal (Exploded Logit) 30% 20% 10% 0% Absolute Error Discussion Several features are of note within the models presented above, particularly in relation to which appear to perform poorly for certain types of data. For the cardinal data, the Proximate method appears to perform relatively poorly. This may be a feature of the data used here. In the MVH study, states were selected so that all individuals will consider two very mild states, three mild states, three intermediate states and three severe states. Here, the rankings between similarly severe states will typically involve very low cardinal differences between the states but potentially large differences in the descriptions of the states. As a result, the coefficients in this model are particularly small. This has the effect of reducing the range of unscaled health estimates, and exaggerating the possible range of health state values when anchored at 0 and 1. We might expect, and do observe, smaller unscaled coefficients and a similar lack of fit for the ordinal Proximate model. (In contrast the SGbased model used in McCabe et al considered states that were more evenly spread, so that the suppression of coefficient values observed here was not a factor.) For the ordinal data, the vs Full Health model has a very high coefficient on dead that is nonsignificant. By the exclusion criteria, being in full health is always preferred to being dead within our sample, so there is unanimity and the logit regression is unable to estimate a standard error for the

14 14 coefficient hence, the lack of significance. The implication of the large (unscaled) coefficient on dead is that the other coefficients also increase in value. The ordinal vs Dead and ordinal vs Pits models also perform relatively poorly in terms of the number of insignificant or omitted variables. Here, we can safely predict that most states will be preferred to being dead, that most states predicted by the EQ-5D will be preferred to the worst state describable by the instrument. It is probably unsurprising therefore that these models are unable to extract as much information from the dataset as others, and consequently have a worse performance. One puzzling finding is the significance of the constant term in the ordinal preference models. Given that the states compared are randomly swapped, there should be no obvious preference for choosing the first or the second state in a comparison. Indeed for the original Exhaustive model, only 49.9% of responses selected the first state, whilst 50.1% of responses selected the second state. However, the constant terms suggest that the first state is typically preferred. Given that none of the independent variables (after switching) had significant means, there is no clear justification for a constant term. Removing the constant term from the regression makes no difference to any of the estimated coefficients to two decimal places. In the constant difference models, the constant term is insignificant except in the case of the Exhaustive model. Rather than suggesting poor performance of the other models (at it is unlikely to have an impact), this may instead suggest a problem with the general method. Here, independence of irrelevant of alternatives is assumed to obtain almost eight times the number of data points as used in other models and this larger sample size allows more items to be considered and may exaggerate significance. We considered using a random ranking-based model like Random 1 or Random 2 as our preferred ordinal preference model for a related concern that the Exhaustive model simply extracts too much data from limited preferences. Overall, the VAS Levels regression outperformed the cardinal differences regression vs Full Health. The VAS Levels regression considered disutility, as a loss against full health. However, the cardinal difference regression appears to do the same thing considering a difference versus full health. The former outperformed the latter for two reasons firstly, taking differences reduces the sample size instead of having n observations from an individual we have only (n-1) observations, and so less information to estimate the underlying regression. The second reason is structural: the constant in the levels regression is informative because it functions as an N2 term (in a similar way to the N3 term). Here, any deviation from full health starts at 0.84 (i.e ) plus whatever additional decrement is suggested by the coefficients. In the cardinal difference regressions, this difference is lost and hence is not measured as this role is significant in the levels regression, the fit of the model is poorer without it, and this is reflected in the results. Conclusions

15 15 This paper has compared alternative analyses for different ordinal preferences over health states and cardinal differences between health states, alongside a regression estimating the levels of health. We have identified preferred models for the ordinal preference and cardinal difference states. In contrast to previous studies we have found that ordinal preference models appear to give different results to cardinal data. The similarly severe states considered in the MVH study may have contributed to this. Unless similarly close states can be avoided (e.g. in experimental design), then exploding a ranked list of states for use in logit models appears unwise. Similarly, we have concerns about the reliance of an exhaustive logit on independence of irrelevant alternatives. McCabe et al found this assumption to be violated, and we have additional concerns over possibly exaggerated significance. Given that the ordinal preference data performs worse than the cardinal difference and levels regression, it seems relatively clear that: (1) VAS contains at least some relevant and useful cardinal data and (2) ignoring such data worsens performance of the resulting measures. A more general issue relates to whether ordinal preference data should be used in valuation tasks. There are a variety of different means by which ordinal preference data could be used and this paper has not considered the full variety of methods available, with DCE and/or Best-Worst both alternative options (see for example, Ryan et al, 2001). However, in standard valuation settings, it is likely that rank data will be elicited by post and based on tasks such as VAS rather than on more intensive SG or TTO data. Both Salomon and McCabe et al used ranking data based on more intensive methods and both found that the ordinal data obtained were adequate. This is not a guarantee that ranking data per se is adequate, and the findings here are to the contrary.

16 16 References Brazier, J. and McCabe, C Is there a case for using visual analogue scale valuations in CUA by Parkin and Devlin a response: Yes there is a case, but what does it add to ordinal data? Health Economics. 16: Dolan, P., Gudex, C., Kind P., Williams, A The measurement and valuation of health: First report on the main survey. Centre for Health Economics, University of York: York Dolan, P., and Roberts, J Modelling Valuations for Eq-5d Health States: An Alternative Model Using Differences in Valuations. Medical Care 40(5): Johannesson, M., Jonsson, B., Karlsson, G Outcome measurement in economic evaluation. Health Economics. 5: McCabe, C., Brazier, J., Gilks, P., Tsuchiya, A., Roberts, J., O Hagan, A., Stevens, K Using rank data to estimate health state utility models. Journal of Health Economics. 25: McCabe, C., Stevens, K., Roberts, J., Brazier, J.E. 2005a. Health state valuations for the Health Utilties Index Mark 2 descriptive system: result from a UK valuable survey. Health Economics. 14: McCabe, C., Stevens, K., Brazier, J.E. 2005b. Utility values for Health Utilties Index Mark 2: An empirical assessment of alternative mapping functions. Medical Care. 43: Parducci, A., and Weddell, D.H The category effect with rating scales: number of categories, number of stimuli, and method of presentation. Journal of Experimental Psychology: Human Perception and Performance 12: Parkin, D. and Devlin, N Is there a case for using visual analogue scale valuations in costutility analysis? Health Economics. 15: Roberts, J. and Dolan, P To what extent do people prefer health states with higher values? A note on evidence from the EQ-5D valuation set. Health Economics. 13:

17 17 Robinson, A., Loomes, G., and Jones-Lee, M Visual Analog Scales, Standard Gambles, and Relative Risk Aversion. Medical Decision Making 21:17-27 Ryan, M., Scott, D.A., Reeves, C., Bate, A., van Teijlingen, E.R., Russell, E.M., Napper, M., Robb, C.M Eliciting public preferences for health care: a systematic review of techniques. Health Technology Assessment; 5(5): Salomon, J.A Reconsidering the use of rankings in the valuation of health states: a model for estimating cardinal values from ordinal data. Population Health Metrics. 1:12. (accessed ) Schwartz, A Rating scales in context. Medical Decision Making 18:236. Torrance G.W., Feeny, D.H., Furlong Visual analogue scales; do they have a role in the measurement of preferences for health states? Medical Decision Making. 21: